Research map / 3D vision / Gaussian splatting
Fast Feedforward 3D Gaussian Splatting Compression
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes FCGS, an optimization-free feed-forward compressor for 3D Gaussian Splatting that compresses a 3DGS scene in a single pass using a multi-path entropy module and inter/intra-Gaussian context models.
Problem
Existing 3DGS compression methods require per-scene finetuning, making compression slow (minutes to hours); the paper targets rapid compression of any existing 3DGS without optimization.
Approach
A feed-forward autoencoder with a multi-path entropy module that adaptively routes color attributes through an autoencoder path or direct quantization via a learned binary mask, geometry attributes are always quantized directly, plus inter-Gaussian context models that build feed-forward grids from previously decoded Gaussians and intra-Gaussian channel-chunk context models, with GMM probability es
Key results
- FCGS achieves over 20x compression on DL3DV-GS (Ours-highrate 27.44 MB at 29.26 dB vs 372.50 MB for 3DGS), with lowrate 15.83 MB at 28.86 dB.
- Encodes about 1 second per 100K Gaussians on a single GPU, with encoding of 1.57M Gaussians in about 16-20 s total.
- Compresses 3DGS from feed-forward models MVSplat and LGM zero-shot with compression ratios of 15x and 5x.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Remove inter- and intra-Gaussian context models (ablation) | FCGS · DL3DV-GS | bit consumption: 1.5× | Hurt |
| Set all color attributes to the autoencoder path (m=1) (ablation) | FCGS · DL3DV-GS | fidelity | Hurt |
| Set all color attributes to the direct-quantization path (m=0) (ablation) | FCGS · DL3DV-GS | storage cost | Hurt |
| Process geometry attributes through the autoencoder instead of directly quantizing them (ablation) | FCGS · DL3DV-GS | ability to render correctly | Hurt |
| Use MEM-inferred masks instead of setting all color masks to m=0 (ablation) | FCGS on feed-forward MVSplat 3DGS · MVSplat | PSNR (dB): approximately 0.7dB | Hurt |
Surprising findings
- Setting all color attributes to the autoencoder path (m=1) causes a large fidelity drop despite no quantization, showing MLP non-invertibility hurts rendering.
- Feeding geometry attributes through the autoencoder makes training collapse because rasterization breaks.
Limitations
- Performance gap versus per-scene optimization methods such as HAC and ContextGS.
- Domain gap for 3DGS from feed-forward models; mask must be forced to all zeros for MVSplat and LGM.
- Not applied to anchor-based 3DGS variants.
Experimental setup
Datasets: DL3DV-GS (derived from DL3DV) MipNeRF360 Tank&Temples Deep Blending ACID Gobjaverse
Benchmarks: DL3DV-GS MipNeRF360 Tank&Temples Deep Blending ACID Gobjaverse
Metrics: PSNR SSIM LPIPS size (MB) encoding time
Baselines: LightGaussian Navaneet (compact 3DGS) Simon (compression with finetuning) SOG EAGLES Lee et al. HAC ContextGS MSC
Builds on: 3DGS hyperprior image compression Scaffold-GS
Closest papers by meaning
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.94)
- ContextGS : Compact 3D Gaussian Splatting with Anchor Level Context Model (NeurIPS 2024 · similarity 0.93)
- Temporal Smoothness-Aware Rate-Distortion Optimized 4D Gaussian Splatting (NeurIPS 2025 · similarity 0.93)
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- Locality-aware Gaussian Compression for Fast and High-quality Rendering (ICLR 2025 · similarity 0.92)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.92)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.